How much is the stability of reanalysis products affected by the increasing number of observations assimilated?

Reanalysis products are becoming an increasingly popular option for monitoring climate due to (1) their comprehensive assessment of Earth processes (providing multiple variables of different domains), (2) their absence of gaps, (3) their global spatial coverage, and (4) their long temporal coverage, with some products extending back to the 1950s. However, reanalysis estimates have larger uncertainties than satellite products and may also present stability issues, so the fitness of each reanalysis variable for monitoring a particular climate application needs to be assessed. Reanalysis datasets combine numerical models and observations with a data assimilation scheme that controls the observations assimilated and their weight in the model. Both uncertainty and stability of reanalysis datasets strongly depend on this assimilation scheme and the number of observations available.  

Producing a data assimilation scheme that optimizes both uncertainty and stability is not straightforward as both properties have opposing requirements. On the one hand, the uncertainty of reanalysis estimates is reduced by increasing the number (and weight) of observations assimilated. This is typically achieved in most recent years due to the abundance of satellite and in-situ observations. On the other hand, the former approach creates a stability challenge. Both the satellite and in-situ observations available exponentially increase in time, so artificial trends and discontinuities can be potentially introduced in the climate data record. The magnitude of these trends/discontinuities depends on the variable evaluated (observations are assimilated only for some variables), the change in the number of observations in time, the spatial region (some observations are unevenly distributed), and the weight given to the observations. All these factors need to be evaluated to determine whether the uncertainty, and particularly the stability, of a reanalysis product is adequate for monitoring a specific climate variable based on the requirements established by GCOS. 

In this study, we evaluate the data assimilation scheme of the most well-known atmospheric reanalyses (ERA5, JRA-55, MERRA-2) and the land component of ERA5 (ERA5-Land). We focus on two variables with different assimilation schemes: snow cover, a variable with direct assimilation of in-situ and satellite observations, and snow albedo, a variable without direct assimilation of observations that depends strongly on the snow data assimilated. The products evaluated also present different assimilation schemes: ERA5 and JRA-55 assimilate an increasing number of snow observations in time, ERA5-Land does not directly assimilate observations but is indirectly forced by ERA5 fields, and MERRA-2 assimilates precipitation observations but not snow observations. We evaluate their fitness for climate monitoring by using in-situ snow depth observations as a reference to quantify their uncertainty and stability. We also inter-compare their global/regional snow cover and snow albedo trends to evaluate their stability globally.